Gemma 4 Developer Agent Competition 2026: $100,000 for offline coding agents and research

The Gemma 4 Developer Agent Competition pays $100,000 in cash for turning a small open model into an offline coding agent that fixes real Python bugs, and for research on how to build one. The main Kaggle leaderboard carries $65,000 and closes on Wednesday 2 December 2026 at 23:59 UTC. An optional Paper Track adds $35,000 for original research and closes first, on Thursday 12 November 2026 at 23:59 UTC. Both moments appear in your local time in the dates table. Entry is free, teams can have up to five people, and the competition is open worldwide apart from a few sanctioned regions and sanctioned persons.

This guide covers both tracks: what you actually build, every deadline in UTC and in your local time, who can enter and from which countries, the exact submission format, how scoring works and the rules that quietly cost people their entries. Every fact comes from the official Kaggle competition pages and settings, the rules, the starter notebook and Kaggle Staff and host posts on the competition forums, checked on 17 October 2026.

New to Kaggle? Read the short answer, then follow the step-by-step section in order. Already training models? Jump to the rules that trip people up and the tips, because the 12-hour time budget is where most of the strategy lives.

Hackathon guide · Google DeepMind (sponsor Google LLC, run by Kaggle)

Gemma 4 Developer Agent Competition

Open

Time left to joinWed 25 Nov 2026, 23:59 UTC

Last moment to join: Wed 25 Nov 2026, 23:59 UTC

Prize pool

$100,000

$65,000 for the main leaderboard plus $35,000 for the optional Paper Track, all cash in US dollars

Deadline (UTC)

Wed 2 Dec 2026, 23:59 UTC

Main track Final Submission Deadline as stated: December 2, 2026, 11:59 PM UTC. Accept the rules by November 25, 2026, 11:59 PM UTC. The Paper Track closes November 12, 2026, 11:59 PM UTC.

Who can enter

Anyone 18 or older (or the local age of majority, if higher) with a Kaggle account, worldwide with a few named exclusions

Excluded: residents of Crimea, DNR, LNR, Cuba, Iran and North Korea, and anyone subject to US export controls or sanctions. Under-18s need parent or guardian consent. Google and Kaggle staff can enter but cannot win.

Format

Online on Kaggle, teams of up to 5

The main track is scored automatically on about 120 hidden Python tasks. The Paper Track is judged on a Kaggle Writeup of up to 3,000 words.

Open the official pageOfficial rulesAll open hackathons

Gemma 4 Developer Agent Competition: the short answer

What is Gemma 4 Developer Agent Competition?

The Gemma 4 Developer Agent Competition is a Kaggle coding agent competition hosted by Google DeepMind under the theme “Accelerate research in autonomous coding agents.” Google LLC is the official sponsor and Kaggle Inc. runs it on Google’s behalf. Kaggle is an online platform for data science and machine learning competitions, where you write and run code in the browser. Gemma 4 is Google’s family of open models, which means the model weights, the learned numbers that make up the model, can be downloaded and run on your own hardware instead of being reached only through a paid service. The full brief sits on the official competition overview.

The main track asks you to post-train Gemma 4 into an autonomous software engineering agent. Post-training means taking a model that has already been trained and improving it further, for example with better prompts, tools or small add-ons made by fine-tuning, which means training an existing model a little more on a smaller set of examples. A coding agent is a program in which the model reads code, runs commands, edits files and repeats until it has a fix. Your agent has to find its way around a real Python code repository and draft a patch, which is the set of line-by-line changes that fixes the problem, on everyday hardware with no internet. Kaggle labels this track a Featured Prediction Competition.

Scoring works like SWE-Bench, a well-known test in which an AI system must fix real software issues and only gets credit if that issue’s validation tests pass afterwards. Your agent faces about 120 hidden Python bug-fix and feature tasks curated from private repositories, split evenly between a public leaderboard, which shows your score while the contest runs, and a private leaderboard, which is the only ranking that decides prizes. The development set you can practise on has 129 public tasks drawn from four popular Python libraries: fastapi, rich, requests and httpx.

The second track is the optional Paper Track, a separate Kaggle competition labelled a Featured Hackathon. Here you write up original, unpublished research on agentic software engineering in a Kaggle Writeup of up to 3,000 words, a project report page that you write and submit directly on the competition site. Suggested topics are Tuning and Optimization (such as parameter-efficient fine-tuning, which trains a small set of extra weights instead of the whole model, and reinforcement learning, where a model learns from rewards such as a test passing), Code Comprehension (code graphs, parsing and embeddings, which are numeric fingerprints of code that let a program find similar snippets), Tasks and Benchmarks, and Graph Reasoning. You do not need to enter the main competition to submit a paper. The official pages say top submissions will be highlighted at a Google hosted event (such as a NeurIPS expo workshop, NeurIPS being a major AI research conference), and as of 17 October 2026 they gave no date or place. On the same day Kaggle showed more than 11,300 entrants and about 3,000 teams on the main track, while the Paper Track showed 4,952 entrants but only 250 participants in 243 teams.

Key dates and deadlines in UTC and your local time

Kaggle sets every deadline at 11:59 PM UTC on the day shown unless the rules say otherwise. UTC (Coordinated Universal Time) is the world’s reference clock and never shifts for daylight saving. The table below gives each date as the rules state it, the same moment in UTC and, worked out automatically by your browser, the time where you are. In the United States every deadline in this competition falls after clocks change on 1 November 2026, so 23:59 UTC is 3:59 pm PST (Pacific) and 6:59 pm EST (Eastern) on the same date. Anywhere ahead of UTC, such as most of continental Europe, Asia and Australia, each deadline lands on the next calendar day: a deadline listed as 2 December closes there on 3 December.

The two tracks run on separate calendars. The Paper Track closes almost three weeks before the main competition, and the main competition has an earlier entry deadline that is easy to miss. If you have not accepted its rules by 25 November, you cannot submit at all.

Treat the main entry deadline as a hard gate. Accepting the rules costs nothing, so do it as soon as you think you might take part, even if you plan to build later. The same goes for teams: once 25 November 2026 at 23:59 UTC passes, nobody can join or merge teams in the main competition. On the Paper Track, Kaggle’s settings close new entries and team mergers at the same moment as final submissions, 12 November at 23:59 UTC.

Do not plan to submit on the final day. You get one main track submission per day, you cannot submit again until the previous one has finished scoring, and each scoring run gives your agent up to 12 hours, including sandbox setup and vLLM start-up, before patch validation even starts. The sandbox is the sealed environment your agent’s code runs in, and vLLM is the software that serves the Gemma 4 model. In early October 2026 Kaggle Staff reported submissions waiting more than 15 hours in the queue, which led to a fix and a pending submission limit. Your last realistic chance to see a fresh score comes several days before 2 December.

The rules give no calendar dates for the steps after the contest. Winners have one week to answer the notification email, two weeks to return the prize documents, and payment follows within about 30 days of those documents arriving. As of 17 October 2026, no date or place had been announced for the Google hosted event that will highlight top submissions.

Prizes

All prizes are cash in US dollars, and the overview lists total prizes available of $100,000. Because the two tracks are separate Kaggle competitions, the money splits into $65,000 for the main leaderboard and $35,000 for the Paper Track. Prizes are paid net of any taxes the sponsor is legally required to withhold, and all taxes on prizes are the winner’s own responsibility, so the amount that finally reaches you depends on the tax rules where you live.

Team prizes are split evenly among eligible members unless the whole team unanimously agrees a different split and tells Kaggle before payout. Prizes cannot be transferred to someone else.

On the Paper Track, a host said one writeup can be considered for several awards but can win only one, and a team can submit at most two writeups. The Paper Track awards no Kaggle points or medals. The main competition does award them, so a strong leaderboard finish helps your public Kaggle profile even if you miss the money.

Who can enter, and from which countries

You need a registered Kaggle.com account, and only one account per person. You must be 18 or the age of majority where you live (the age at which you legally become an adult), whichever is older, unless the sponsor agrees and a parent or guardian gives consent. A pinned Kaggle Staff post on the main competition forum says under-18s can take part by submitting two parent or guardian consent forms, one for using kaggle.com and one for each competition, and that children under 13 are not allowed on Kaggle at all.

The competition is open to residents of the United States and worldwide except residents of Crimea, the so-called Donetsk People’s Republic (DNR) and Luhansk People’s Republic (LNR), Cuba, Iran and North Korea, and anyone subject to US export controls or sanctions. In practice, if you do not live in one of those named regions and are not a sanctioned person, the country rule does not stop you. If a winner lives in a country where prizes cannot be awarded, that person is not eligible for a prize. The competition rules place no restriction on students versus professionals, and Kaggle Staff confirmed that individuals with no company affiliation can receive prizes.

Employees, interns, contractors, officers and directors of Google LLC, Kaggle Inc. and their parents, subsidiaries and affiliates may enter, but they cannot win prizes.

Teams and solo entries

Both tracks allow teams of up to five people. Each member registers individually on Kaggle first and then confirms membership through the team notification. You can belong to only one team per competition.

In the main competition the team leader performs mergers, but the merged team’s total submission count must not exceed the maximum allowed as of the Team Merger Deadline, which is the daily limit multiplied by the number of days the competition has run. The last moment to join or merge is Wednesday 25 November 2026 at 23:59 UTC, shown in your local time in the dates table.

On the Paper Track, Kaggle’s settings close team mergers on Thursday 12 November 2026 at 23:59 UTC. A host clarified that a team can submit at most two writeups, and the judges decide which award each one is eligible for.

Sharing rules are strict. Private sharing of code or data outside your team is not allowed, and sharing code between teams counts as a breach unless the teams merge. Any code you share publicly must go on this competition’s Kaggle forums or notebooks, and by sharing it you license it under an open source licence that allows commercial use.

How to enter Gemma 4 Developer Agent Competition, step by step

Everything happens inside Kaggle, so you need only a browser to start, wherever you are. A Kaggle notebook is a free, browser-based coding environment that runs on Kaggle’s servers, with GPUs (graphics processors that run AI models quickly) attached. You write Python in cells, run them, and the submission.zip the notebook writes to its output folder (/kaggle/working) is what gets scored.

These steps follow the official overview, the Data tab and the Kaggle Staff starter notebook. Steps 1 to 9 are the main track. Step 10 is the optional Paper Track, which you can do on its own.

What you must submit

The main track asks for a single zip file, not a video, a pitch or a slide deck. Nobody reads your code to rank you: the only thing that counts is how many hidden tasks your agent fixes. The Paper Track is the opposite, a written research report judged by people.

Your agent gets 12 hours to submit patches for all tasks. That clock includes sandbox setup and, Kaggle Staff confirmed, the vLLM server start-up, but not patch validation. Public and private tasks run in the same 12-hour run, so you cannot spend extra time on the tasks that decide prizes.

The main track needs no video, public repository, demo link or written description to enter. If you win, though, you must deliver your training code, inference code, a description of the compute environment and reproducible documentation, so keep your work tidy from the start.

On the Paper Track you must actually click Submit: drafts or unsubmitted Writeups at the deadline are not considered. A host said posting your paper to arXiv is fine and you can even submit the arXiv link. Any private Kaggle resource you attach to the Writeup becomes public after the deadline, and any linked notebook must be readable without a login or paywall. The rules do not state a required language for either track.

How entries are judged

The two tracks are judged in completely different ways. The main track is scored by a machine with a PASS/FAIL test on each task, and only the hidden private half decides the prizes. The Paper Track is scored by judges against a five-part rubric.

The Public Leaderboard shows how your agent does on one half of the roughly 120 hidden tasks while the competition runs. The Private Leaderboard uses the other half and is the only ranking that decides prizes. According to the Data tab, every test task was checked against a larger frontier model (one of the most capable AI models available), which either passed it or came within a single test case of passing.

Tools, APIs and what they cost

You must use Gemma 4, specifically one of the four QAT variants available on Kaggle Models. Agents are written as Google ADK agent configs and compiled into ADK agents. The harness uses the swegemma, adk-submission and adk-eval-core libraries and serves the model with vLLM, with max_model_len set to 32768 in the starter notebook. That number is the context window: the most tokens (word pieces) the model can see at once, so long files and noisy test output eat into it fast.

The competition has access to Kaggle’s L4x4 machines, which carry 96 GB of GPU memory in total. They are restricted to notebooks attached to this competition (trying to get around that risks a ban), must run with internet disabled, and use your GPU quota at twice the rate of T4x2 or P100 machines (other Kaggle GPU machine types), a rate the overview says Kaggle may raise if needed. The weekly size of that quota is not stated on the competition pages. On money, the rules say “NO PURCHASE NECESSARY TO ENTER OR WIN”, Kaggle accounts and notebooks are free, and no cash credits are offered.

External data and models are allowed if they are publicly available and equally accessible at no cost to all participants, or if they meet the host’s Reasonableness Standard: a small subscription such as Gemini Advanced is acceptable, but a proprietary dataset licence costing more than a prize is not. AutoML tools (software that builds models automatically) are allowed with an appropriate licence, and any open source code you use must be under an OSI-approved licence (one recognised by the Open Source Initiative) that does not limit commercial use. For the Paper Track, a host said coding assistants such as Codex are permitted, but following their terms of service is your responsibility.

Rules that trip people up

These are the rules and platform limits most likely to cost you an entry, taken from the rules and from Kaggle Staff and host posts on the competition forums.

Ownership, prize payment and tax

You warrant that your submission is your own original work and that you own it, and the rules do not otherwise transfer ownership of non-winning entries. Winning is different. The rules list the winner licence type as Open Source Apache 2.0, a permissive licence that lets anyone use, modify and build on the code, including commercially. If you win, you license your winning submission and the source code used to generate it under an OSI-approved licence that in no event limits commercial use, with carve-outs for commercially available third-party software and for data or pretrained models with incompatible licences. You must also deliver the final model’s training code, inference code and a description of the compute environment with documentation, and you may be asked for a detailed, reproducible write-up that links to a code repository. The same licence terms apply to the Paper Track.

Data access and use is “Competition Use and Commercial Apache 2.0”, so you may use the Competition Data for any commercial or non-commercial purpose, but you must not pass it on to anyone who has not accepted the rules. Google and Kaggle may use your name and likeness for promotion, and Kaggle Staff said reasonable accommodations around likeness are possible for winners.

Potential winners are notified by email and must respond within one week. Within two weeks of being asked, they sign and return the prize acceptance documents: eligibility certifications, licences and releases, and a US tax form. That is IRS Form W-9 for US residents, or IRS Form W-8BEN for everyone else, a form in which a non-US person confirms their foreign tax status. Prizes are paid within about 30 days after Google or Kaggle receives those documents, net of any taxes the sponsor must withhold, and all taxes are the winner’s sole responsibility. The rules do not state the payment method or any withholding rate, so check with a tax adviser in the country where you live.

Project ideas that fit the brief

These are our own suggestions, not ideas from Google or Kaggle. Each one stays inside the rules: one Gemma 4 model for every agent, only harness tools and agent_tool sub-agents, no internet at run time, and a submission.zip within the starter notebook’s 3 GiB check.

Tips to give your entry the best chance

• Our advice: do the time arithmetic first. Twelve hours across about 120 hidden tasks works out to about 6 minutes per task if they run one after another, and that budget also has to cover sandbox setup and vLLM start-up. A bigger model that reasons better but runs out of time on half the tasks can score worse than a smaller, faster one, so test both on the development set.
• Use eval_config.yaml to set hard per-task limits such as max_time_minutes and max_tool_calls, so one stubborn task cannot eat the time the other tasks need.
• Treat each daily submission as precious. Run the starter notebook’s Phase 2 verification on the 129 public development tasks before you submit, change one thing at a time and keep a simple log of what each submission tested.
• Accept the rules for both tracks as soon as you are even half interested. It is free, accepting the main track rules unlocks the dataset, and it protects you from the 25 November entry gate on the main track.
• Check every deadline in your own time zone, not just in UTC. Each one closes at 23:59 UTC, which is the afternoon or evening of the same day in the Americas but already the next calendar day across most of Europe, Asia and Australia, so put the local time from the dates table in your calendar.
• Read the pinned posts and Kaggle Staff replies in each competition’s Discussion tab, because that is where rule clarifications appear. Kaggle Staff said the official Kaggle Discord is not monitored by staff or hosts, so put important questions on the forums.
• Keep your work reproducible from day one: clean training and inference code, a note of your compute setup and only OSI-licensed open source code. Winners must hand all of this over, and rebuilding it after the deadline is painful.
• For the Paper Track, write the paper yourself. A host warned that purely LLM-generated writing will not score well, and Verifiability is a fifth of the score, so link a public notebook that reproduces your results and click Submit well before 12 November.

Please note: we update this page from time to time, but hackathon dates, prizes and rules can change quickly, and there can be a gap between a change and our update. Check the official rules before you enter or submit. This guide helps you plan; it is not legal or tax advice, and ReconScribe is not responsible for decisions made on it.

Gemma 4 Developer Agent Competition frequently asked questions

What is the Gemma 4 Developer Agent Competition?

It is a Gemma 4 Kaggle competition hosted by Google DeepMind, with Google LLC as sponsor, in which you post-train an open Gemma 4 model into an offline coding agent. The agent is scored on about 120 hidden Python bug-fix and feature tasks in a SWE-Bench style test. A separate, optional Paper Track rewards original research on agentic software engineering.

Which countries can enter the Gemma 4 Developer Agent Competition?

The rules open it to residents of the United States and worldwide, except residents of Crimea, the so-called Donetsk People’s Republic (DNR) and Luhansk People’s Republic (LNR), Cuba, Iran and North Korea, and anyone subject to US export controls or sanctions. You must also be 18 or the age of majority where you live, whichever is older, or use Kaggle’s parent or guardian consent process. A winner living in a country where prizes cannot be awarded is not eligible for a prize, and winners outside the US sign IRS Form W-8BEN to be paid.

When is the Gemma 4 Developer Agent Competition deadline?

The main competition’s final submission deadline is Wednesday 2 December 2026 at 23:59 UTC. You must accept the rules and finish any team merger by Wednesday 25 November 2026 at 23:59 UTC, and the Paper Track closes on Thursday 12 November 2026 at 23:59 UTC. In the United States each of these is 3:59 pm PST or 6:59 pm EST on the same date, while anywhere ahead of UTC, such as most of continental Europe, Asia and Australia, it falls on the next calendar day. The dates table on this page shows each one in your local time.

How much prize money is there?

There is $100,000 in cash in total. The main track pays $37,000, $18,000 and $10,000 to the top three on the Private Leaderboard, and the Paper Track pays $15,000 for Overall Best Paper and $10,000 each for Best New Resource and Best New Application. Team prizes are split evenly unless the whole team agrees otherwise and tells Kaggle before payout.

Does it cost anything, and do I need my own GPU?

No. The rules say no purchase is necessary, and Kaggle accounts and notebooks are free. Notebooks attached to this competition can use Kaggle’s L4x4 machines with 96 GB of GPU memory, but they use your GPU quota at twice the normal rate and must run with internet disabled. The size of the weekly quota is not stated on the competition pages.

Which Gemma 4 models are allowed?

Four quantized variants are allowed: gemma-4-31b-it-qat-w4a16-ct, gemma-4-12b-it-qat-w4a16-ct, gemma-4-e4b-it-qat-w4a16-ct and gemma-4-e2b-it-qat-w4a16-ct. You pick exactly one, and every agent and sub-agent in your submission must use it, although each agent can use a different LoRA adapter. Kaggle Staff said the smaller variants are scored on the same hardware with the same 12-hour limit.

Do I have to enter the main competition to join the Gemma 4 Paper Track?

No. The Paper Track is a separate Kaggle competition with its own rules, and you can enter it on its own. Joining one track does not enter you in the other, so accept the rules for each one you want to enter.

How many times can I submit?

The main track allows 1 submission per day, and you must wait for your previous submission to finish scoring before you submit again. You can select up to 2 Final Submissions for judging, and Kaggle selects them for you if you do not. The Paper Track allows up to 5 submissions per day and up to 2 Final Submissions, and a host said a team can submit at most two writeups.

Can I use Claude, GPT or other AI tools?

At run time your agent can use only the one Gemma 4 model you chose, with no internet, but while you build, external data and models are allowed if they are publicly available at no cost to everyone or meet the host’s Reasonableness Standard, under which a small subscription such as Gemini Advanced is acceptable. On the Paper Track a host said coding assistants such as Codex may be used to write code, that following their terms of service is your responsibility, and that purely LLM-generated writing will not score well. As of 17 October 2026 there was no official ruling on fine-tuning Gemma 4 with trajectories generated by Claude or GPT.

What happens if my agent goes over the 12-hour limit?

The 12 hours cover all hidden tasks, including sandbox setup and vLLM start-up, and public and private tasks run in the same run. Participants report that going over the limit errors the whole submission, but the overview does not say this. Kaggle Staff said in early October 2026 that a request to score unfinished tasks as 0 instead was being reconsidered, so check the Discussion tab for updates.

Can students or under-18s take part?

The rules place no restriction on students versus professionals. Under-18s can take part by submitting two parent or guardian consent forms, one for using Kaggle and one for the competition, according to a pinned Kaggle Staff post on the main forum. Children under 13 are not allowed on Kaggle.

Do winners have to open-source their solution?

Yes. The winner licence type is Open Source Apache 2.0, and winners must license the winning submission and its source code under an OSI-approved licence that does not limit commercial use. Winners also deliver training code, inference code, a description of the compute environment and reproducible documentation.

More open hackathons

Bolt Builders Week hackathon$17,500Enter by Wed 21 Oct 2026, 06:59 UTCMeta Global AI Developer Hackathon$1,000,000Enter by Mon 26 Oct 2026, 03:59 UTCNebius x NVIDIA Global AI Hackathon$50,000 cash plus 4 NVIDIA Jetson Orin Nano kitsEnter by Fri 30 Oct 2026, 17:00 UTCPayPal AI Hackathon$67,500 cashEnter by Thu 12 Nov 2026, 20:00 UTCMeta VR Start Developer Competition 2026Up to $1,000,000Enter by Wed 18 Nov 2026, 20:00 UTCKiroween 2026$100,000 in cashEnter by Fri 4 Dec 2026, 22:00 UTCNFL Big Data Bowl 2027$100,000Enter by Wed 6 Jan 2027, 23:59 UTCName That LLM: Identify the Model Behind the Text$50,000Enter by Wed 6 Jan 2027, 23:59 UTC

See every open hackathon, with filters and deadlines in your local time

Related reading

• Hackathons 2026: every open hackathon, deadlines in your local time
• Bolt.new free tokens and the Bolt hackathon
• Meta VR Start Developer Competition: complete guide
• Meta AI Developer Hackathon: complete guide
• Kiroween 2026: complete guide
• Free AI API credits and free tiers
• AI leaderboard: every major model ranked
• Claude Code tutorial for building faster
• Free credits and tools for students

Sources

• Gemma 4 Developer Agent Competition: overview, timeline, prizes and evaluation
• Gemma 4 Developer Agent Competition: rules
• Gemma 4 Developer Agent Competition: Data tab
• Gemma 4 Developer Agent Paper Track: overview
• Gemma 4 Developer Agent Paper Track: rules
• Official Getting Started notebook (Kaggle Staff)
• Pinned Kaggle Staff post: how to get started
• Pinned Kaggle Staff post: new Gemma 4 models
• Pinned Kaggle Staff post: Kaggle Error and queue delays
• Kaggle Staff answers to aggregated questions
• Paper Track host answer: writeups per team
• Paper Track host answer: arXiv preprints
• Paper Track host and staff answers: AI assistance and individual winners